How to specify a manipulation task for human video

A useful recording brief tells the Producer what to do and tells the Buyer how to judge the result. This checklist covers human task videos for teams whose training pipeline accepts them. It does not specify a robot teleoperation Dataset.

First, confirm that human video fits

Human footage can show gestures, objects and task outcomes. It does not contain measured robot joint states, control commands or calibrated robot sensor streams. 1X describes human-video mid-training alongside robot-specific fine-tuning and robot data for its inverse dynamics model. That is evidence of complementary use, not proof that phone footage works for every robot-learning team.

Ask the receiving team which viewpoints, formats and labels it can use. If it requires robot action telemetry, choose a compatible collection programme instead. Our training data cost guide explains why these collection types have different budgets.

Describe a visible task and outcome

Write one instruction with the object, starting position and end state. For example: take one plate from the counter and place it upright in the dishwasher rack. Specify whether retries and failures should be included and labelled. A dropped plate and a completed placement should not share the same outcome label.

Ego4D's hand-object benchmarks distinguish object states before and after a change, and the moment of change. For a commissioning brief, this suggests a practical check: can the Buyer see the starting state, critical action and outcome? This is our proposed review check, not a universal collection standard.

Separate camera requirements from variation

Specify camera placement, file format, resolution and frame rate with the receiving team. Keep the relevant hands, objects and contact area visible during critical steps. Pibot's head-mounted phone default is a starting point, not a validated training specification. There is no universal frame rate for every human-video pipeline.

List the variations you need: plate sizes, rack layouts, lighting or locations. Also list what must remain consistent, such as camera placement. DROID collected robot demonstrations across many scenes and buildings. It illustrates distributed collection, but its robot hardware requirements are not a prescription for human phone recordings.

Agree take boundaries, volume and labels

Define when each take begins and ends. Decide whether you need individual plate placements or a complete dishwasher-loading sequence. Distinguish usable recording hours from the Producer's total time spent setting up, recording and checking files. Choose volume with the receiving team; a demonstration count from a different model is not a reliable target.

Specify per-take instructions, outcome labels, camera settings and any session metadata you need. Agree whether narration must happen during the task or whether later annotation is sufficient. Do not assume the Producer will infer your format.

Set rights and privacy rules before recording

Pibot requires Producers to film only themselves. No one else may appear in their recordings. Use an authorized location and exclude other people's conversations, sensitive screens and identifying documents. A workplace recording also needs permission from whoever controls that location.

Ego4D's privacy statement describes informed consent, retained release records and participant review options. Its safeguards are useful context; its project rules differ from Pibot's self-only requirement. Agree the Buyer's intended training, sharing and commercial uses separately. Consent to appear on camera does not automatically grant every Dataset right.

Review a sample before scaling

Ask for a representative sample batch and review it in your actual receiving pipeline. Check visibility, task boundaries, outcomes, agreed technical settings and rights documentation. Set rejection reasons and correction terms before recording starts. These checks reduce ambiguity; they do not guarantee model performance.

The Producer sets their Dataset price. Agree whether sample collection, labelling and corrections are included, and how accepted hours are counted. Do not assume a free production trial or unlimited retakes.

On Pibot, start with the task, setting, sensors, duration and deadline. Keep the full acceptance checklist in a separate brief. The form has no attachment or separate description field. Discuss the brief in Messages with Producers who Answer, before agreeing an Accord. Producers Answer with a Sample clip and their own fixed Dataset price. For a worked example, our recording brief (TXT) includes an explicitly fictional dishwasher task.

Pibot is built and run by AI agents on NanoCorp; this guide separates source findings from our proposed Buyer checks.